shithub: libvpx

Info  •  Files  •  Log  •  Branches

ppc: d63 predictor 16x16

About 18x faster.

Change-Id: Id043bf76c011e03e992085bb5e20f330d3e98cd4

--- a/test/test_intra_pred_speed.cc

+++ b/test/test_intra_pred_speed.cc

@@ -320,7 +320,8 @@

                 vpx_dc_left_predictor_16x16_vsx, vpx_dc_top_predictor_16x16_vsx,

                 vpx_dc_128_predictor_16x16_vsx, vpx_v_predictor_16x16_vsx,

                 vpx_h_predictor_16x16_vsx, vpx_d45_predictor_16x16_vsx, NULL,

-                NULL, NULL, NULL, NULL, vpx_tm_predictor_16x16_vsx)

+                NULL, NULL, NULL, vpx_d63_predictor_16x16_vsx,

+                vpx_tm_predictor_16x16_vsx)

 INTRA_PRED_TEST(VSX, TestIntraPred32, vpx_dc_predictor_32x32_vsx,

                 vpx_dc_left_predictor_32x32_vsx, vpx_dc_top_predictor_32x32_vsx,

--- a/vpx_dsp/ppc/intrapred_vsx.c

+++ b/vpx_dsp/ppc/intrapred_vsx.c

@@ -541,3 +541,24 @@

     row1 = vec_perm(row1, above_right, sl1);

   }

 }

+

+void vpx_d63_predictor_16x16_vsx(uint8_t *dst, ptrdiff_t stride,

+                                 const uint8_t *above, const uint8_t *left) {

+  const uint8x16_t a0 = vec_vsx_ld(0, above);

+  const uint8x16_t a1 = vec_vsx_ld(16, above);

+  const uint8x16_t above_right = vec_splat(a1, 0);

+  const uint8x16_t b = vec_perm(a0, above_right, sl1);

+  const uint8x16_t c = vec_perm(b, above_right, sl1);

+  uint8x16_t row0 = vec_avg(a0, b);

+  uint8x16_t row1 = avg3(a0, b, c);

+  int i;

+  (void)left;

+

+  for (i = 0; i < 8; i++) {

+    vec_vsx_st(row0, 0, dst);

+    vec_vsx_st(row1, 0, dst + stride);

+    dst += stride * 2;

+    row0 = vec_perm(row0, above_right, sl1);

+    row1 = vec_perm(row1, above_right, sl1);

+  }

+}

--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl

+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl

@@ -116,7 +116,7 @@

 specialize qw/vpx_d45_predictor_16x16 neon ssse3 vsx/;

 add_proto qw/void vpx_d63_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";

-specialize qw/vpx_d63_predictor_16x16 ssse3/;

+specialize qw/vpx_d63_predictor_16x16 ssse3 vsx/;

 add_proto qw/void vpx_h_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";

 specialize qw/vpx_h_predictor_16x16 neon dspr2 msa sse2 vsx/;